Chemical park-oriented VOC gas leakage robot intelligent inspection system and method

Abstract

This application relates to the field of chemical testing and discloses a robotic intelligent inspection system for VOC gas leaks in chemical industrial parks. The system includes: multiple mobile inspection robotic agents for collecting VOC concentration data and environmental parameters at their locations; and a processing unit for: constructing or updating a physical model describing the VOC diffusion process based on the VOC concentration data and environmental parameters; generating or updating a model uncertainty field that quantifies the uncertainty of the physical model at various points in space based on the deviation between the prediction results of the physical model and the VOC concentration data; and generating the next action command for the multiple mobile inspection robotic agents based on the VOC concentration and the model uncertainty field. By introducing and quantifying a model uncertainty field parallel to the VOC concentration field, and using the information entropy value in this field as a key basis for generating decision commands, the efficiency and purposefulness of robotic inspection are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of chemical testing, specifically to a robotic intelligent inspection system and method for VOC gas leaks in chemical industrial parks. Background Technology

[0002] Volatile organic compounds (VOCs) are common raw materials, intermediates, or products found in chemical industrial parks. Because these compounds are typically flammable, explosive, and exhibit varying degrees of biotoxicity, accidental leaks during production, storage, or transportation pose a serious threat to personnel safety, the ecological environment, and production facilities. Therefore, real-time, accurate, and large-scale monitoring of VOC concentrations within chemical industrial parks, along with rapid response and location of leaks, is a core element in ensuring safe production in these parks. In recent years, combining mobile robot technology with gas sensing technology to utilize robots to replace or assist manual environmental inspections has become an important technological development direction in this field, finding applications in smart parks, industrial IoT, and environmental emergency response scenarios.

[0003] However, existing robot-based gas inspection technologies still have a series of inherent limitations when applied to complex chemical industrial park scenarios. Current robot inspection technologies mostly employ preset fixed paths or climbing strategies based on concentration gradients. The former is slow to respond to dynamically changing leak events and lacks flexibility. The latter easily leads to robots getting stuck at local concentration extremes or performing unnecessary back-and-forth movements in already sufficiently detected areas, resulting in low detection efficiency. When deploying multiple robots, existing solutions typically lack effective coordination mechanisms. Each robot, as an independent detection unit, lacks global-level optimization in its action decisions, easily leading to overlapping detection areas or omissions of critical areas. More critically, existing inspection methods generally lack physical modeling of the gas diffusion process. This prevents the system from predicting the leak situation based on physical laws, leaving inspection behavior at a passive, "post-event" response level. Therefore, after confirming a leak, existing technologies struggle to provide crucial source-tracing information such as the specific location of the leak source and the leak rate, and also cannot accurately monitor the diffusion range dynamically, resulting in significant deficiencies in accident assessment and control capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a robotic intelligent inspection system and method for VOC gas leaks in chemical industrial parks. It solves the problems of rigid robot inspection strategies, low detection efficiency, lack of effective collaboration among multiple agents, and inability to accurately identify leak sources and dynamically warn of risk areas due to a lack of physical understanding of gas diffusion processes in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a robotic intelligent inspection system for VOC gas leaks in chemical industrial parks. The system includes multiple mobile inspection robotic agents and a processing unit. The multiple mobile inspection robotic agents are used to collect VOC concentration data and environmental parameters characterizing the physical state of their respective locations.

[0006] The processing unit is communicatively connected to the plurality of mobile inspection robot agents, and is specifically used to perform the following operations: Based on the VOC concentration data and environmental parameters received from the multiple agents, a physical model describing the diffusion process of VOCs in the environment is constructed or updated. Based on the deviation between the prediction results of the physical model and the VOC concentration data, a model uncertainty field is generated or updated. This field is used to quantify the degree of uncertainty of the physical model at various spatial locations. Based on the known VOC concentration and the model uncertainty field, the next action command for the multiple mobile inspection robot agents is generated.

[0007] Through the above technical solution, this system no longer relies solely on concentration gradients for inspection, but introduces the concepts of physical models and model uncertainty. The generation of action commands depends simultaneously on the currently known leak situation and the uncertainty of environmental perception, giving the robot's inspection behavior a dual objective: tracking leaks and optimizing the model. This mechanism establishes a feedback loop: the agent collects data by executing commands, the data is used to reduce model uncertainty, and a more accurate model can generate more efficient commands, thereby improving the overall efficiency and accuracy of the system in responding to leak events.

[0008] In an optional embodiment, the model uncertainty field is quantified using information entropy. Regions with large model prediction bias or sparse environmental data are assigned higher information entropy values, and vice versa.

[0009] In an optional embodiment, the processing unit generates the action command by constructing a multi-objective dynamic reward function. This function includes a concentration detection reward and a model optimization reward. The concentration detection reward incentivizes the agent to move to high-concentration regions, while the model optimization reward incentivizes the agent to move to high-information-entropy regions.

[0010] In one optional embodiment, the numerical value of the model optimization gain is positively correlated with the uncertainty value at the corresponding location in the model's uncertainty field. This allows the agent to be directly guided to the key location that best enhances the system's overall cognitive level for data collection.

[0011] In an optional embodiment, the processing unit employs a distributed Q-learning algorithm for decision-making. Each agent autonomously calculates and selects action instructions that maximize its long-term expected return based on the multi-objective dynamic reward function, thus achieving distributed collaboration among multiple agents.

[0012] In an optional embodiment, the physical model is specifically a reaction-diffusion partial differential equation model used to describe the diffusion, convection, and chemical reaction processes of VOCs.

[0013] In an optional embodiment, the processing unit is further configured to determine whether the global uncertainty of the model uncertainty field is lower than a preset convergence threshold. When it is lower than the threshold, it indicates that the physical model has reached a high confidence level. At this time, the processing unit will determine the specific location, leakage rate, and start time of the VOC leakage source through collaborative inversion calculation based on the high confidence physical model and all collected VOC concentration data.

[0014] In an optional embodiment, after determining the VOC leakage source, the processing unit is further configured to: predict the future VOC diffusion envelope based on the high-confidence physical model, and generate new action instructions to assign the multiple mobile inspection robot agents to move to several key nodes of the diffusion envelope, thereby constructing a dynamic virtual monitoring barrier to monitor changes in the leakage range.

[0015] In an optional embodiment, the multi-objective dynamic reward function further includes a forward-looking early warning reward. This reward is generated by the processing unit based on the physical model's prediction of the potential formation of high-concentration or high-uncertainty regions in the future, and is used to guide the agent in forward-looking location deployment.

[0016] A second aspect of this invention provides a robotic intelligent inspection method for VOC gas leaks in chemical industrial parks. The method includes the following steps: Multiple mobile inspection robots deployed within the chemical industrial park collect VOC concentration data and environmental parameters at their respective locations. Based on the collected VOC concentration data and environmental parameters, a physical model describing the diffusion process of VOCs in the environment is constructed or updated. Based on the deviation between the prediction results of the physical model and the VOC concentration data, a model uncertainty field is generated or updated. This field is used to quantify the degree of uncertainty of the physical model at various spatial locations. Based on the known VOC concentration and the model uncertainty field, the next action command of the multiple mobile inspection robot agents is generated and executed.

[0017] This invention provides a robotic intelligent inspection system and method for VOC gas leaks in chemical industrial parks. It offers the following advantages: 1. This invention constructs and identifies an online physical model of the reaction-diffusion partial differential equation describing the VOC diffusion process, and uses this model to predict future gas diffusion trends, thus enabling the inspection system to have foresight. The robot's action commands are no longer based solely on the currently detected concentration, but rather on predictions of potential high-concentration or high-uncertainty areas in the future. This achieves a shift from passive response to proactive early warning and forward-looking deployment, improving the timeliness of response to sudden leaks.

[0018] 2. This invention introduces and quantifies a model uncertainty field parallel to the VOC concentration field, and uses the information entropy value in this field as the key basis for generating decision instructions, greatly improving the efficiency and purposefulness of robot inspection. The agent's actions are given a clear objective for model optimization, namely, actively going to areas that can minimize the uncertainty of the system's global cognition to conduct detection, avoiding ineffective repeated inspections in areas with sufficient information, and accelerating the system's convergence speed to the true state of the environment.

[0019] 3. This invention achieves efficient collaborative operation among multiple mobile inspection robot agents by employing a distributed Q-learning algorithm based on a multi-objective dynamic reward function. Each agent can make decisions autonomously and in a distributed manner, balancing multiple objectives such as concentration detection and model exploration within a unified reward framework, and implicitly considering the behavior of other agents. This naturally facilitates division of labor and cooperation, avoids action conflicts and data collection redundancy, and maximizes the global information acquisition efficiency of the entire robot swarm.

[0020] 4. By setting a convergence threshold for the uncertainty field of the model, this invention can trigger collaborative inversion calculations after the physical model reaches a high level of confidence. This not only detects the existence of leaks but also accurately inverts key source parameters such as the location of the leak source, the leakage rate, and the start time. This ability to trace the cause from the observation of a phenomenon provides direct and crucial decision support information for locating the root cause of an accident, defining responsibility, and subsequent precise containment.

[0021] 5. This invention achieves a resource-optimized and highly adaptable risk monitoring mode by accurately locating the leakage source, using a converged high-precision physical model to predict the future VOC diffusion envelope, and assigning multiple agents to construct dynamic virtual monitoring barriers at key nodes of the envelope. Compared to fixed monitoring points or aimless patrols, this dynamic barrier can adaptively adjust with changes in the leakage situation, achieving the most effective monitoring and alarm of the risk boundary with minimal resources. Attached Figure Description

[0022] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a schematic diagram of the data receiving and fusion module according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the physical model construction and identification module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the model uncertainty quantification module in an embodiment of the present invention; Figure 5 This is a schematic diagram of the decision instruction generation module according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example: Please see the appendix Figure 1 -Appendix Figure 5 This invention provides a robotic intelligent inspection system for VOC gas leaks in chemical industrial parks, comprising: Mobile inspection robot intelligent agent: In this embodiment, the mobile inspection robot is a physical unit deployed within the chemical industrial park to perform data acquisition and mobility tasks. It is not an isolated detection device, but rather a key, mobile sensing and execution node in the closed-loop process of the intelligent inspection system described in this invention.

[0025] The mobile inspection robot intelligent agent integrates a perception module, a positioning and navigation module, a communication module, and an execution module in its structure. Each module works in concert to achieve its specific function in the method of the present invention.

[0026] Specifically, the perception module is an integrated multi-sensor unit, whose core purpose is to acquire multi-dimensional physical state information of the robot's current location.

[0027] In a preferred embodiment, the sensing module includes a VOC gas sensor for directly measuring the concentration of volatile organic compounds in the air. This measurement is denoted as... This is the foundation for all subsequent calculations in this system; it constitutes the ground truth value for comparison with the physical model's predictions. The processing unit minimizes the difference between the model predictions and this measured value. The residual between, i.e. This allows for online identification and optimization of physical models.

[0028] Furthermore, the sensing module also includes a three-dimensional wind field sensor. This sensor is used to measure the wind speed and direction at the location of the intelligent agent in real time, and its output wind field vector... This is a key environmental parameter for constructing and solving the reaction-diffusion partial differential equation. Specifically, this wind field vector is directly used as the convection term in the equation. Accurate wind field measurements are crucial for the physical model to accurately describe the macroscopic transport process of VOC gases.

[0029] Preferably, the sensing module may further include temperature and air pressure sensors. The environmental parameters collected by these sensors can be used to further correct the diffusion coefficient in the physical model. Since the gas diffusion rate is related to the temperature and pressure of the medium, introducing these parameters can make the physical model more closely resemble the real physicochemical process, thereby improving the accuracy of the entire system's understanding of the environmental state.

[0030] The positioning and navigation module is designed to provide the intelligent agent with high-precision spatiotemporal coordinates.

[0031] In this embodiment, the module can determine the agent's own three-dimensional spatial position in a preset global coordinate system in real time. And record the current timestamp. The spacetime coordinates Physical data collected by the sensing module ( , This data is bound together to form a complete data packet. This ensures that each piece of data uploaded to the processing unit has a clear spatiotemporal index, which is essential for constructing the global VOC concentration field. and model uncertainty field This forms the basis of such spatiotemporal data structures.

[0032] Simultaneously, this module is also responsible for parsing the action instructions issued by the processing unit. The instruction is then converted into a specific navigation path and motion control command to guide the execution module to complete the movement of the physical location.

[0033] The communication module serves as a bridge for bidirectional data exchange between the intelligent agent and the processing unit.

[0034] On one hand, this module is responsible for uploading the multi-dimensional data packets collected by the sensing module and stamped with spatiotemporal times by the positioning module to the processing unit via a wireless communication network. This uplink data stream is the sole source of information driving model identification, uncertainty quantification, and decision-making within the processing unit.

[0035] On the other hand, this module is responsible for receiving the next action instruction for the agent, calculated by the decision instruction generation module of the processing unit. This downlink instruction serves as the basis for all actions of the agent, embodying the global collaborative optimization strategy described in this invention.

[0036] The execution module is the ultimate implementer of the intelligent agent's physical actions. Typically, this module is a robotic chassis with autonomous mobility, such as a wheeled or tracked platform. Based on control commands received from the positioning and navigation module, it drives the robot to move to locations specified by the action commands. The specified target location.

[0037] In summary, the mobile inspection robot agent in this embodiment is defined as an "active information probe." It not only passively collects data, but its data collection behavior (i.e., where to collect data) is guided by a complex algorithm (i.e., a distributed Q-learning algorithm driven by information entropy) designed to reduce global uncertainty in the system. Through the coordinated operation of its perception, localization, communication, and execution capabilities, the agent constitutes a crucial closed loop connecting algorithmic decision-making and physical world information acquisition in the technical solution described in this invention, thereby achieving adaptive and self-optimizing operation of the entire inspection system from data collection to intelligent decision-making.

[0038] Data reception and fusion module: In this embodiment, the data receiving and fusion module constitutes the data entry point of the processing unit. Its function is to act as a bridge connecting multiple mobile inspection robot intelligent agents in the physical world with the abstract computing model inside the processing unit, and to transform the raw and scattered measurement data into structured information that can be used by subsequent advanced algorithms.

[0039] This data receiving and fusion module maintains continuous or intermittent communication with all mobile inspection robot agents deployed within the chemical industrial park through a preferred wireless communication network. Its core function is to receive, process, and organize heterogeneous data sent by each agent.

[0040] Specifically, the functionality of this module includes the following processing flow: Data packet reception and parsing: The module continuously monitors the communication channel and receives asynchronous data packets from different agents in real time. Each data packet encapsulates complete field information of a single agent at a specific time and location.

[0041] This module first parses the received data packet, extracting several key information fields encapsulated within it. These fields include at least: the measured value of the volatile organic compound concentration obtained from the VOC gas sensor. One or more environmental parameters obtained from a three-dimensional wind field sensor, particularly including the wind field vector. ; and the three-dimensional spatial coordinates of the agent provided by the positioning and navigation module, synchronized with the above measurements. and precise timestamps .

[0042] Data validation and structuring: After parsing, the module will perform preliminary verification of the data, checking its integrity and format correctness, and removing obviously abnormal or corrupted data packets.

[0043] Subsequently, the core task of this module is to effectively organize and structure these data points, originating from different agents and discretely distributed across the timeline. It transforms each valid data packet into a unified-format, multi-dimensional spatiotemporal data record and stores it in a centralized real-time database or in-memory data structure. This data record explicitly includes physical measurement values ​​(…). , ) and their spatiotemporal context ( The data was strongly bound together. This step is crucial because it transforms the raw, unordered communication data stream into a unified dataset with spatiotemporal indexes that can be queried and analyzed.

[0044] Provide data support for subsequent modules: The structured dataset processed by this module serves as the direct data foundation for the operation of all other computing modules within the processing unit, providing an indispensable input for the subsequent core algorithm flow of this invention.

[0045] On one hand, this module provides ground truth for the physics model building and identification modules. Specifically, in the identification module, which attempts to minimize the loss function... To optimize the physical model parameters At that time, all the measured concentration values ​​required in this formula And environmental parameters at the corresponding spatiotemporal points (used to calculate the wind field vector of the convection term). All of these are directly derived from the structured dataset provided by this module.

[0046] On the other hand, this module provides a quantitative basis for the model uncertainty quantification module. The uncertainty field... The calculation, its formula In addition to the model prediction error, another key parameter is needed: data density. This module calculates and provides data density information near any given location by statistically analyzing the distribution of received data points in the spatiotemporal dimension, thus supporting the accurate generation of uncertain fields.

[0047] In addition, the module aggregates measured VOC concentration values. The set also constitutes a global VOC concentration field. The discrete data points used for updating or rendering.

[0048] In summary, the data receiving and fusion module in this embodiment goes far beyond simple information collection. By receiving, parsing, verifying, and structuring asynchronous and heterogeneous data from multiple mobile intelligent agents, it constructs a unified, ordered data foundation with complete spatiotemporal context. This module ensures that subsequent complex calculations, including physical modeling, uncertainty quantification, and decision generation, can be performed based on a coherent and reliable data source. It is a fundamental and necessary technical link in realizing the entire intelligent inspection closed-loop process described in this invention.

[0049] Physical model construction and identification module: In this embodiment, the physical model construction and identification module is the core of the processing unit. Its function is to establish a physical model that can mathematically describe the spatiotemporal evolution of VOC gas in a complex environment, and to continuously identify and optimize the model online by using the measured data collected by the mobile inspection robot agent.

[0050] The operation of this module is the key to realizing the transformation of this invention from simple concentration-based inspection to intelligent inspection based on a deep understanding of environmental physical processes.

[0051] In one specific embodiment of the present invention, in order to comprehensively describe the various physical phenomena in the gas diffusion process, the physical model constructed by this module is preferably an improved reaction-diffusion partial differential equation. This equation can unify the molecular diffusion of VOCs, convective transport caused by wind fields, and possible chemical reactions or attenuation processes within a mathematical framework.

[0052] The mathematical expression for this model is as follows: ; In this equation, each mathematical term has a definite physical meaning. The left-hand side of the equation... Item, indicating spatial location and time VOC concentration Rate of change over time.

[0053] The first term on the right side of the equation, This is the diffusion term. This term describes the molecular diffusion of VOCs due to their concentration gradient. It is worth noting that the diffusion coefficient is... It is not a fixed constant, but one of the key parameters that this module needs to identify, and its value is affected by local temperature and air pressure environmental factors.

[0054] The second term on the right side of the equation, This is the convection term. This term describes the macroscopic movement of VOC gas as a whole with the ambient wind field. The wind field vector in this term... The value is not calculated by this module, but is a known input variable whose value comes from the wind field data provided by the data receiving and fusion module and measured in real time by each mobile inspection robot agent.

[0055] The third term on the right side of the equation, This term represents the source / sink term. It describes VOC leakage sources in the environment (as a positive source term) or concentration decay due to photolysis, chemical reactions, or surface adsorption (as a negative sink term). The specific functional form and related parameters of this term are also one of the core objectives that this module needs to identify online.

[0056] The diffusion coefficient is one of the unknown parameters to be identified in the physical model. Relevant parameters and source / sink The relevant parameters together form a parameter set. In the initial stage of system operation, this parameter set is assigned a set of prior values. .

[0057] To enable the aforementioned prior physical model to converge from its initial state to an accurate description of the real physical processes, this module executes a continuous online parameter identification process. This process is essentially an optimization process, aiming to continuously adjust the parameter set. The goal is to find the optimal parameters that minimize the difference between the model's predictions and real-world observations.

[0058] Specifically, this module minimizes a loss function. This enables online parameter identification. The loss function is defined as the model's predicted concentration at all valid observation points. Concentration measured by robot The sum of squared errors between: ; In this loss function, The data is provided by the data receiving and fusion module at a specific time and space point. The ground truth value, and This means that at the same point in time and space, the current parameter set is used. The model prediction values ​​are obtained by solving the aforementioned partial differential equations.

[0059] Upon receiving new observation data, this module employs a preferred optimization algorithm, either gradient descent or Extended Kalman Filter (EKF), to minimize the loss function. The solution process iteratively updates the parameter set. This continuous process of identification and optimization enables the physical model to dynamically and adaptively reflect changes in the environment, and its prediction accuracy improves with the increase of effective data.

[0060] The optimized physical model output by this module is the foundation for all subsequent advanced functions of this invention. It not only provides the basis for the model uncertainty quantification module to calculate prediction errors, but also provides the decision instruction generation module with the predictive capability to calculate forward-looking early warning benefits. After the system's cognitive convergence, it becomes the fundamental tool for performing advanced tasks such as collaborative inversion of leak source location and construction of dynamic virtual monitoring barriers.

[0061] Model uncertainty quantification module: In this embodiment, the model uncertainty quantification module is a core functional unit within the processing unit, running immediately after the physical model construction and identification module. The purpose of this module is not to directly process VOC concentration information, but rather to establish and maintain a digital field—the model uncertainty field—that runs parallel to the physical world and characterizes the system's own cognitive state.

[0062] The necessity of this module lies in its introduction of a completely new decision-making dimension into the intelligent inspection system described in this invention. By quantifying and utilizing "uncertainty," the behavior of the mobile inspection robot agent is no longer merely a passive response to detected high-concentration areas, but rather an active and strategic exploration of unknown or cognitively ambiguous areas, thereby accelerating the system's global understanding of the environmental situation.

[0063] Specifically, the functionality of this module includes the following processing flow: This module is responsible for generating and dynamically updating a global model uncertainty field, denoted as . This field is related to the VOC concentration field. In terms of structure, the corresponding three-dimensional spatiotemporal data field, at any point within the field The value represents the reliability or certainty of the prediction results of the physical model at that point in time and space.

[0064] In a preferred embodiment, this uncertainty is quantified using the concept of information entropy. That is to say, As an information entropy field, the higher the information entropy value of a point in the field, the greater the uncertainty of the system's cognition of the state at that point, and the higher the potential information value contained at that point that can be used to optimize the model.

[0065] In this embodiment, the uncertainty or information entropy value at any given location is not determined by a single factor, but rather by the combined effect of two core factors. This design aims to comprehensively reflect the two main sources of model inaccuracies: errors in the model itself and insufficient data used for modeling.

[0066] The first factor is model prediction bias. When a mobile inspection robot agent collects new measured data at a certain location... Then, the physical model construction and identification module will generate a model prediction value for that point. The absolute deviation between these two This directly reflects the accuracy of the current model in describing the physical state at that point. The greater the deviation, the worse the applicability of the model in that region, and therefore the higher the uncertainty at that point.

[0067] The second factor is local data sparsity. In areas that the robot has not yet visited or has visited very little, even if the physical model can provide predictions, these predictions are extrapolated from data in other areas, and their reliability is inherently low. To quantify this uncertainty caused by the lack of field data validation, this module introduces a data density parameter. This parameter is used to characterize the location point. The frequency or density of data collected in the vicinity over a past period. The lower the data density, the more likely the area is a cognitive "blind spot," and the higher its uncertainty.

[0068] To unify the two factors mentioned above within a single quantitative framework, this module employs a comprehensive calculation function to generate the information entropy value. Its mathematical expression can be defined as: ; In this expression, It is a pre-defined, monotonically increasing composite function. Its characteristic is that, regardless of whether the problem stems from large model prediction bias, low data density, or both, the function will output a high information entropy value that is positively correlated with the actual situation. In this way, the module can map the accuracy information of the physical model and the completeness information of the data acquisition onto this unified uncertainty field.

[0069] The final output of this module is this dynamically updated, complete model uncertainty field. This field is a key input to the subsequent decision-making process of this invention.

[0070] Decision instruction generation module: In this embodiment, the decision instruction generation module is the embodiment of the will and action center of the processing unit. Its core function is to transform the system's abstract cognition of the environment (i.e., the VOC concentration field and the model uncertainty field) into specific physical actions that the multiple mobile inspection robot agents will perform.

[0071] This module is a key link in realizing the transformation from "cognition" to "action" in the intelligent inspection closed-loop process described in this invention. The decision-making algorithm it executes determines the efficiency, purposefulness, and coordination of the robot's inspection behavior, and is the fundamental guarantee of the system's intelligence.

[0072] The overall operating logic of this module can be further broken down into two closely linked steps: first, a quantitative assessment of multi-objective benefits is conducted, and then distributed action decisions are made based on this assessment.

[0073] In one specific embodiment of the present invention, in order to guide the robot’s behavior to achieve multiple task objectives, rather than just tracking concentration, this module first constructs and calculates a multi-objective dynamic benefit function for each potential action (i.e., moving to a candidate target point) of each mobile inspection robot agent.

[0074] This function aims to quantify the immediate, comprehensive value that performing a certain action can bring. Its preferred mathematical expression is: ; In this expression, Representing intelligent agents The total revenue. It consists of multiple weighted revenue components and one cost item. The specific physical meaning and calculation source of each component are as follows: The benefits of concentration detection Its value is positively correlated with the VOC concentration value of the candidate target point. This value comes directly from the global VOC concentration field maintained by the processing unit. This benefit reflects the system's "utilization" of existing knowledge, incentivizing the agent to conduct focused investigations of known, high-risk leakage areas.

[0075] The benefits of exploring the unknown area Its value is negatively correlated with the historical access frequency or data freshness of candidate target points. This benefit term is used to encourage agents to achieve more comprehensive coverage of the entire monitoring area and avoid ineffective back-and-forth movements between a few high-concentration points.

[0076] The model optimization benefits This is a core revenue component of the present invention. Its value is related to the candidate target point in the uncertainty field of the model. The corresponding information entropy value is positively correlated. This uncertainty field is generated by the model uncertainty quantification module. The purpose of introducing this benefit term is to transform the abstract system-level goal of "reducing model uncertainty" into a concrete action incentive for individual robots, thereby driving the robots to actively go to the location that best improves the overall cognitive level of the system for data collection.

[0077] The aforementioned forward-looking early warning benefits This is an optional advanced benefit. Its value comes from the physical model construction and identification module. This module can use the optimized physical model to predict the VOC diffusion trend in the near future, and assign a positive benefit in advance to areas where the current concentration is not high, but may become high concentration or highly uncertain in the future, thereby guiding the agent to make forward-looking deployments.

[0078] The aforementioned action costs It is a negative benefit term whose value is positively correlated with the time, energy consumption, or path risk required to move from the current location to the candidate target point, and is used to ensure the economy and feasibility of the decision.

[0079] After calculating the total reward value for all possible actions, this module needs to solve the action coordination problem among multiple agents to avoid all agents going to the same point with the highest total reward, thus avoiding resource waste.

[0080] To achieve efficient distributed collaboration, this module preferably employs a game-theoretic decision-making algorithm based on distributed Q-learning. Within this algorithm framework, each mobile inspection robot agent is considered an independent decision-making entity, aiming to maximize its own long-term cumulative expected return.

[0081] Specifically, each intelligent agent Each maintains an independent Q-function, denoted as Used to evaluate its state Take action The value of this function is determined by its continuous iterative updates based on the following rules: ; In this updated rule, This refers to the calculations above, which determine the agent's actions. The total immediate benefit obtained afterward. The item represents its arrival at a new state. Then, it is the present value of the maximum expected return that can be obtained in the future. By combining immediate returns with expected future returns, this algorithm enables the agent's decision-making to have a long-term perspective.

[0082] At each decision point, this module calculates the Q-value for each agent for all feasible actions and selects the action that maximizes the Q-value. This serves as the final action command. The command is then encapsulated and sent via the communication module to the corresponding mobile inspection robot agent.

[0083] In summary, the decision command generation module in this embodiment successfully decomposes complex system-level objectives into action guidance for individual robots by constructing a multi-objective reward function that integrates concentration tracking, unknown exploration, model optimization, and forward-looking early warning, and embedding this function into a distributed Q-learning game-theoretic decision-making framework. It not only provides each robot with the "motivation" (reward function) for action but also offers a set of "intelligence" (Q-learning) for collaborative action, thereby ensuring that the entire robot swarm can complete the intelligent inspection task of VOC leak incidents in chemical industrial parks in an efficient, collaborative, and continuously self-optimizing manner.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A robotic intelligent inspection system for VOC gas leaks in chemical industrial parks, characterized in that, include: Multiple mobile inspection robot agents are used to collect VOC concentration data and environmental parameters at their locations; Processing unit, used for: Based on the VOC concentration data and environmental parameters, construct or update a physical model describing the VOC diffusion process; Based on the deviation between the prediction results of the physical model and the VOC concentration data, a model uncertainty field that quantifies the uncertainty of the physical model at various points in space is generated or updated. Based on the VOC concentration and the model uncertainty field, the next action command for the multiple mobile inspection robot agents is generated.

2. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 1, characterized in that, The uncertainty field of the model is quantified by information entropy, and regions with sparse data or large model prediction bias have high information entropy.

3. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 1, characterized in that, The processing unit generates the action command by constructing a multi-objective dynamic benefit function that includes concentration detection benefit and model optimization benefit.

4. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 3, characterized in that, The optimization benefit of the model is positively correlated with the uncertainty value at the corresponding position in the uncertainty field of the model.

5. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 3, characterized in that, The processing unit employs a distributed Q-learning algorithm to make decisions based on the multi-objective dynamic reward function in order to generate the action instructions.

6. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 1, characterized in that, The physical model is a reaction-diffusion partial differential equation model.

7. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 1, characterized in that, The processing unit is further configured to: when the global uncertainty of the model uncertainty field is lower than a preset convergence threshold, calculate the location, leakage rate and start time of the VOC leakage source by collaborative inversion based on the physical model and the VOC concentration data.

8. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 7, characterized in that, The processing unit is further configured to: after calculating the VOC leakage source, predict the future VOC diffusion envelope based on the physical model, and generate instructions to assign the multiple mobile inspection robot agents to build dynamic virtual monitoring barriers at key nodes of the diffusion envelope.

9. The intelligent robotic inspection system for VOC gas leaks in chemical industrial parks according to claim 3, characterized in that, The multi-objective dynamic benefit function also includes a forward-looking early warning benefit, which is generated based on the physical model's prediction of the potential formation of high-concentration or high-uncertainty regions in the future.

10. A method based on the system of claim 1, characterized in that, Includes the following steps: Multiple mobile inspection robot agents collect VOC concentration data and environmental parameters at their locations; Based on the VOC concentration data and environmental parameters, construct or update a physical model describing the VOC diffusion process; Based on the deviation between the prediction results of the physical model and the VOC concentration data, a model uncertainty field that quantifies the uncertainty of the physical model at various points in space is generated or updated. Based on the VOC concentration and the model uncertainty field, the next action command of the multiple mobile inspection robot agents is generated and executed.

Images